gengar-atropos-environments

To build an RL environment for Atropos training, wire the required BaseEnv lifecycle methods—setup, get_next_item, format_prompt, compute_reward, evaluate, and wandb_log—into the multi-turn agent loop to support tool calling and custom reward functions.

Your compute_reward function fails during agentic evaluation if it does not score rollouts using AgentResult.messages or lacks tool-based verification via ToolContext within the evaluate method's GengarLoop execution.

The correct way to implement CLI workflows for Atropos environments is by configuring the three modes—serve, process, and evaluate—with provider-agnostic inference setup prompts and correct flags for deterministic evaluation.

Yes, you can use wandb logging to track metrics in your Atropos RL environment by implementing the wandb_log method and calling super to ensure metrics integrate properly with the Gengar multi-turn agent loop.

GengarLoop handles tool-based verification by running full agentic rollouts with tools included, using ToolContext sandbox verification to validate agent actions before computing rewards from AgentResult.messages.

Yes, you need to implement all BaseEnv methods—setup, get_next_item, format_prompt, compute_reward, evaluate, and wandb_log—to run an Atropos environment in evaluate mode and maintain the Gengar agent loop contract.